MPR-CNN-based temperature stress extraction method and system
By constructing and training the MPR-CNN network model, the problem of signal-to-noise ratio deterioration in long-distance sensing of the BOTDA system is solved, and fast and accurate temperature stress detection is achieved, which improves the robustness and real-time monitoring capabilities of the system.
Patent Information
- Application Number
- CN202510484241.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-17
- Publication Date
- 2025-08-08
AI Technical Summary
The deterioration of the signal-to-noise ratio of the tail end optical signal during long-distance temperature and stress sensing leads to serious measurement errors, and traditional methods increase costs or affect real-time monitoring capabilities.
The temperature stress extraction method based on MPR-CNN is adopted. By constructing and training the MPR-CNN network model, the MPR module is used for noise reduction processing, and the CNN module is used for rapid data processing, so as to extract the temperature and stress information of the BGS image.
It realizes fast and accurate detection of long-distance temperature stress, reduces computing resource consumption, enhances the robustness and generalization capabilities of the model, and adapts to different application environments.
Smart Images
Figure CN120449119A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of sensing technology, and in particular to a temperature stress extraction method and system based on MPR-CNN. Background Art
[0002] Optical fiber offers advantages such as low cost, immunity to electromagnetic interference, and low loss. Distributed Brillouin Optical Time Domain Analysis (BOTDA) systems using it as a sensing medium can achieve sensing distances on the order of hundreds of kilometers. However, when applied to long-distance temperature and stress sensing, a degraded signal-to-noise ratio (SNR) at the optical end can significantly increase sensing errors, leading to severe measurement errors. Furthermore, the spatial resolution of a BOTDA system depends on both signal strength and noise levels. A degraded SNR reduces the system's ability to distinguish adjacent measurement points, resulting in a decrease in spatial resolution. To improve the SNR, data acquisition time may need to be increased, perhaps through signal averaging. This can increase the measurement cycle time and compromise real-time monitoring capabilities.
[0003] When BOTDA optical signals propagate through optical fibers, losses cause the signal-to-noise ratio (SNR) at the optical end to decrease. This is expressed as: P(z) = P0exp(-αz), where P(z) represents the detected optical power at point z, P0 represents the input optical power, and α represents the optical fiber loss coefficient. For single-mode fiber, α is approximately 0.2 dB / km. This equation shows that when the sensing distance exceeds 50 km, the optical power at the fiber end drops by at least 10 dB compared to the input power, representing significant losses. To achieve ultra-long-distance BOTDA temperature and stress sensing, this intrinsic loss must be compensated or suppressed. Traditional approaches, which incorporate relay equipment such as optical amplifiers into the optical sensing link, often significantly increase system costs or reduce system versatility. Therefore, using advanced signal processing methods to reduce system resource overhead during data processing has become a more popular research direction. Summary of the Invention
[0004] The purpose of the present invention is to provide a temperature stress extraction method and system based on MPR-CNN to achieve long-distance temperature stress detection.
[0005] To solve the above technical problems, the present invention provides a temperature stress extraction method based on MPR-CNN, comprising: S1. Build and train the MPR-CNN network model based on the sensor device; S2. Obtain BGS images through sensing devices and input them into the MPR-CNN network model to obtain temperature and stress information.
[0006] According to the above scheme, the MPR-CNN network model includes an MPR module and a CNN module; The MPR module consists of an encoder-decoder subnetwork and an original resolution processing layer. The encoder-decoder subnetwork extracts the global context information of the input BGS image, and the original resolution processing layer performs convolution on the original resolution of the input BGS image to restore the fine details of the input BGS image. The encoder-decoder subnetwork is horizontally connected to the original resolution processing layer. The MPR module fuses the global context information and fine details to output a denoised BGS image.
[0007] According to the above scheme, the CNN module includes an input layer, a convolution layer, a pooling layer, and an output layer. The input layer is a one-dimensional vector, and the size of the input layer matches the number of BGS sweeps in the BGS image. The number of convolution layers, the size of the convolution kernel, and the size of the pooling layer are all set based on the complexity of the Brillouin gain spectrum of a single sensor point in the BGS image. The output layer does not use an activation function or uses a linear activation function.
[0008] According to the above scheme, the training process of the MPR module includes MPR training data construction and MPR training optimization. The MPR training data construction includes: constructing three-dimensional BGS training data with scanning frequency, fiber sensing distance, and signal amplitude as X, Y, and Z axes respectively; converting the three-dimensional BGS training data into grayscale images in the XY two-dimensional plane; adding Gaussian white noise to the grayscale image to generate noisy training data with different signal-to-noise ratios; and matching the noisy training data with corresponding noise-free grayscale images as true labels.
[0009] According to the above scheme, the MPR module training optimization includes: inputting noisy training data into the MPR module; calculating the loss based on the denoised BGS image output by the MPR module and the true label of the noisy training data; backpropagating to update the MPR module parameters, and iterating until the loss converges.
[0010] According to the above scheme, the loss function used in the loss calculation of the MPR module is the mean square error or the structural similarity index.
[0011] According to the above scheme, the training process of the CNN module includes CNN training data construction and CNN training optimization. The CNN training data construction includes: using Brillouin frequency shift as a training label, and generating CNN training data consisting of multiple BGS images according to the set Brillouin frequency shift variation range, Brillouin linewidth variation, and signal-to-noise ratio; and normalizing or standardizing the CNN training data so that the CNN training data conforms to the standard normal distribution.
[0012] According to the above scheme, CNN training optimization includes: inputting CNN training data into the CNN module; calculating the loss based on the temperature and stress prediction information output by the CNN module and the training labels of the CNN training data; backpropagating to update the CNN module parameters, and iterating until the loss converges.
[0013] According to the above scheme, CNN training optimization includes: using the previous CNN module as the starting point of the current CNN training module for transfer learning; and applying the previous CNN module to the previous sensor system.
[0014] The present invention also provides a temperature stress extraction system based on MPR-CNN, comprising: A sensing device, used for detecting and acquiring a BGS image; The MPR-CNN network model is used to obtain temperature and stress information based on BGS images.
[0015] Beneficial effects By adopting the MPR module, the present invention achieves effective noise reduction processing for BGS images with low signal-to-noise ratios, providing a foundation for subsequent regression tasks. By adopting the CNN module, it achieves rapid processing of large amounts of BGS image data. It is suitable for sensing devices for long-distance detection and adapts to different application environments and conditions, with good robustness and generalization performance.
[0016] Furthermore, the MPR module extracts global contextual information through an encoder-decoder subnetwork and restores fine details in combination with the original resolution processing layer. The two are horizontally connected to achieve feature fusion. This can restore rich contextual information while preserving key spatial details of the BGS image, effectively suppress noise while avoiding signal distortion, and provide high-fidelity denoised BGS images for the subsequent CNN module, fundamentally improving the basic data quality for temperature and stress extraction.
[0017] Furthermore, the CNN module designs the input layer and convolution layer based on the one-dimensional frequency domain characteristics of a single BGS, dynamically adjusts the network parameters according to the complexity of the signal, avoids the redundant calculations of traditional two-dimensional processing, and significantly improves the efficiency of extracting frequency domain features. While ensuring the accuracy of temperature and stress regression, it can achieve a processing speed of tens of thousands of BGS data in seconds, which is several orders of magnitude faster than traditional methods.
[0018] Furthermore, by adopting transfer learning technology, the CNN module trained in the previous sensing system is used as a starting point for fine-tuning, which allows rapid adaptation to new scenarios without training from scratch, significantly reducing computing resource consumption and training time, and enhancing the model's generalization ability for different sensing devices. This allows this solution to be efficiently applied to a variety of long-distance detection scenarios, further improving engineering practicality and robustness. BRIEF DESCRIPTION OF THE DRAWINGS
[0019] Figure 1 This is a flow chart of the temperature stress extraction method based on MPR-CNN in Example 1 of the present invention; Figure 2 : This is the Brillouin gain spectrum measured by the BOTDA system of Example 2 of the present invention; (a) corresponds to SNR = 20 dB, and (b) corresponds to SNR = 5 dB; Figure 3 Schematic diagram of the MPR-CNN structure of the second embodiment of the present invention. DETAILED DESCRIPTION
[0020] To make the purpose, technical solutions, and advantages of the embodiments of the present disclosure more clear, the technical solutions of the embodiments of the present disclosure will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present disclosure. Obviously, the described embodiments are part of the embodiments of the present disclosure, not all of the embodiments. Based on the described embodiments of the present disclosure, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present disclosure.
[0021] Example 1: See also Figure 1 This embodiment discloses a temperature stress extraction method based on MPR-CNN, including: S1. Build and train the MPR-CNN network model based on the sensor device; S2. Obtain BGS images through sensing devices and input them into the MPR-CNN network model to obtain temperature and stress information.
[0022] Furthermore, the MPR-CNN network model includes an MPR module and a CNN module; The MPR module consists of an encoder-decoder subnetwork and an original resolution processing layer. The encoder-decoder subnetwork extracts the global context information of the input BGS image, and the original resolution processing layer performs convolution on the original resolution of the input BGS image to restore the fine details of the input BGS image. The encoder-decoder subnetwork is horizontally connected to the original resolution processing layer. The MPR module fuses the global context information and fine details to output a denoised BGS image.
[0023] Furthermore, the CNN module includes an input layer, a convolution layer, a pooling layer, and an output layer; the input layer is a one-dimensional vector, the size of the input layer matches the number of BGS sweeps in the BGS image, the number of convolution layers, the size of the convolution kernel, and the size of the pooling layer are all set based on the complexity of the Brillouin gain spectrum of a single sensor point in the BGS image, and the output layer does not use an activation function or uses a linear activation function.
[0024] Furthermore, the training process of the MPR module includes MPR training data construction and MPR training optimization. The MPR training data construction includes: constructing three-dimensional BGS training data with scanning frequency, fiber sensing distance, and signal amplitude as X, Y, and Z axes respectively; converting the three-dimensional BGS training data into grayscale images in the XY two-dimensional plane; adding Gaussian white noise to the grayscale image to generate noisy training data with different signal-to-noise ratios; and matching the noisy training data with corresponding noise-free grayscale images as true labels.
[0025] Furthermore, the MPR module training optimization includes: inputting noisy training data into the MPR module; calculating the loss based on the denoised BGS image output by the MPR module and the true label of the noisy training data; backpropagating to update the MPR module parameters, and iterating until the loss converges.
[0026] Furthermore, the loss function used in the loss calculation of the MPR module is the mean square error or the structural similarity index.
[0027] Furthermore, the training process of the CNN module includes CNN training data construction and CNN training optimization; CNN training data construction includes: using Brillouin frequency shift as a training label, generating CNN training data consisting of multiple BGS images according to the set Brillouin frequency shift variation range, Brillouin linewidth variation, and signal-to-noise ratio; and normalizing or standardizing the CNN training data so that the CNN training data conforms to the standard normal distribution.
[0028] Furthermore, CNN training optimization includes: inputting CNN training data into the CNN module; calculating the loss based on the temperature and stress prediction information output by the CNN module and the training labels of the CNN training data; updating the CNN module parameters by backpropagation, and iterating until the loss converges.
[0029] Furthermore, the CNN training optimization includes: performing transfer learning by using the previous CNN module as the starting point of the current CNN training module; and applying the previous CNN module to the previous sensor system.
[0030] This embodiment also provides a temperature stress extraction system based on MPR-CNN, including: A sensing device, used for detecting and acquiring a BGS image; The MPR-CNN network model is used to obtain temperature and stress information based on BGS images.
[0031] Example 2: This embodiment discloses an MPR-CNN network model that acquires the required temperature and stress information based on BGS data measured by a BOTDA system. The MPR-CNN model comprises an MPR module and a CNN module. To accurately extract temperature and stress over long distances, the MPR-CNN network model described in this embodiment must meet the following requirements: The MPR module adopts a multi-stage architecture, with each stage responsible for restoring different aspects of the image during training, such as contextual information and spatial details. Training requires ensuring that each stage works collaboratively to improve restoration results. During training, feature fusion between different stages should be regulated to ensure that contextual information from earlier stages is effectively transferred to later stages, enhancing the expressive power of features. The MPR module can incorporate a supervised attention module, leveraging supervisory signals from real images (high signal-to-noise ratio BGS images) to guide feature refinement and improve the targetedness and effectiveness of feature transfer. At each stage of training, a pixel-by-pixel adaptive design is employed to reweight local features using in-situ supervised attention, enhancing the model's ability to process local details. In addition to designing an effective information exchange mechanism between each stage, lateral connections between feature processing blocks must be addressed to prevent information loss.
[0032] Data preprocessing during CNN module training should ensure that the input data is properly normalized or standardized so that its distribution approaches the standard normal distribution, which facilitates model convergence. For BGS temperature and stress regression tasks, data augmentation can help the model learn more robust features. The CNN network structure can be designed to effectively extract useful features from the input data, such as by selecting appropriate convolutional layers, pooling layers, and activation functions. When training the CNN module, the CNN network's tolerance for the input signal-to-noise ratio should be maximized while ensuring that temperature and strain errors meet requirements. This helps conserve training resources for the MPR module and enables simpler network training and combination.
[0033] Based on the above requirements, the MPR module training process provided in this embodiment is as follows: The multi-stage progressive image restoration extraction convolutional neural network needs to be properly trained to obtain the optimal parameters in temperature and strain extraction to improve the noise reduction effect. Since the signal-to-noise ratio often drops below 5dB in the actual application of long-distance, non-relay BOTDA systems, it is necessary to consider most possible situations during the training of the MPR (multi-stage processing) module. The training set is constructed by combining images, with the scanning frequency as the X-axis, the fiber optic sensing distance as the Y-axis, and the amplitude received by the oscilloscope as the Z-axis to form a three-dimensional BGS (Brillouin gain spectrum) graph. During training, the XY plane image is used as the input data (such as Figure 2 (a) shows the low signal-to-noise ratio BGS image of the distance; Figure 2 (a) It can be seen that the BGS has a clear Lorentz line shape. The peak frequency information of the gain spectrum can be obtained by traditional curve fitting to infer the strain value at that point in the fiber under test. Figure 2 (b) shows a single BGS envelope at low signal-to-noise ratio. The Brillouin gain spectra can be stacked up over distance to obtain Figure 3 The figure shows a three-dimensional BGS (Brillouin gain spectrum) graph, with the scanning frequency as the X-axis, the fiber sensing distance as the Y-axis, and the amplitude received by the oscilloscope as the Z-axis. The training set is constructed using an image combination approach, with XY plane images used as input data. It can be seen that when stress is applied to a certain point in the fiber, the BGS gain peak position changes. This change can be learned as an image feature by the MPR-CHN network, thereby enabling strain extraction. Normalization of the training set converts each image into a grayscale image, facilitating rapid access by the network model. For example, with a scanning frequency range of 10.751 GHz–11.150 GHz, a sensing distance of 50 km, and a spatial resolution of 2 m, the image size is 400 × 25,000. This data needs to be generated through simulation. Furthermore, to simulate noisy BGS, additive white Gaussian noise of varying intensities can be added to the generated BGS images to obtain BGS images with varying signal-to-noise ratios. During network training, the images used to construct the training set must be labeled with corresponding real-world image labels. Since the training images are distance-frequency graphs, the features required for denoising rely on the area surrounding the gain spectrum peak at each measurement point. By assigning a specific, corresponding, noise-free BGS label to each image, the correct strain label can be assigned to each training image during supervised learning. This helps the MPR module learn the mapping between image pairs and achieve high-fidelity, information-free data denoising. The MPR framework consists of three stages, progressively restoring the image. The first two stages, based on encoder-decoder subnetworks, learn extensive contextual information, while the final stage operates at the original input image resolution, preserving the fine textures required in the final output image.
[0034] The CNN module training process provided in this embodiment is as follows: During training, it should be noted that the CNN extraction process is for the temperature extraction of a single BGS at each sensing point (e.g. Figure 2 (As shown in the single Brillouin gain spectrum in [1]), its input layer should be a one-dimensional vector whose size should match the number of BGS sweeps, that is, 400×1. Due to its structure and training process, regression tasks typically use mean squared error (MSE) or mean absolute error (MAE) as the loss function. Depending on the specific task requirements, other loss functions may also be considered, such as Huber loss or mean absolute percentage error (MAPE).
[0035] The network training set needs to consider the variation range of strain or temperature, so Brillouin frequency shift is selected as the training label. Through this process, the temperature and stress can be extracted separately. Therefore, the range of Brillouin frequency shift variation Δv is set to -10MHz to 80MHz, with a step size of 1MHz; the range of Brillouin linewidth variation is set to 55MHz to 75MHz, with a step size of 1MHz; the range of signal-to-noise ratio variation is set to 5dB–20dB, with a step size of 5dB. The total number of BGS is 91×21×4=7644. After training, the network can accurately recover the Brillouin frequency shift information in a single BGS, that is, Figure 3 The convolutional neural network module shown ( Figure 3 The paper demonstrates the network structure and flow chart for obtaining the stress distribution along the optical fiber by processing the Brillouin gain spectrum data obtained from BOTDA measurements using MPR. Convolution kernels of various sizes are stacked to form encoder and decoder structures, preventing information loss while reducing the number of parameters. After high-fidelity denoising using the MPR module, the clean BGS is input into the CNN to obtain the temperature or stress distribution along the optical fiber. The high-fidelity denoised BGS, restored by a multi-stage image restoration network, is then input into the module to obtain the corresponding temperature and stress-distance maps, enabling accurate sensor information extraction. Ultimately, the root mean square error for temperature extraction is less than 1°C, and the error for stress extraction is less than 50 microstrain.
[0036] Based on the complexity of a single BGS, the network depth and width should be adjusted. For example, the number of convolutional layers, the size of the convolution kernel, and the size of the pooling layers all require fine-tuning. Networks that are too deep or too wide may lead to overfitting, while networks that are too shallow or too narrow may fail to capture complex features. Regularization techniques such as dropout and weight decay (L2 regularization) should be applied to regularize the training data to reduce the risk of overfitting, and batch normalization should be used to accelerate training and improve model generalization. The network activation function is also a key technical consideration. In the output layer, activation functions are typically omitted or used that are suitable for regression tasks, such as linear activation functions. In hidden layers, the ReLU function or its variants are commonly used. Network reuse for different sensor systems can utilize transfer learning techniques. Using a pretrained model as a starting point for the feature extractor, the model can be fine-tuned on related tasks to improve training efficiency and performance. Furthermore, due to the MPR denoising module, no or minimal addition of noisy BGS data is required during CNN training to improve robustness, further reducing computational resources. By addressing these technical issues, the performance and generalization ability of the MPR-CNN model in temperature and stress extraction tasks can be improved.
[0037] It can be understood that the above-mentioned MPR-CNN network model is a modular network. The modular network itself is an idea of network structure design, which generally refers to a network model containing two or more sub-modules. These sub-modules learn parameters in their respective information spaces without affecting each other.
[0038] To verify the effectiveness of the MPR-CNN network model provided in this embodiment, the following experimental verification methods can be selected: For the long-distance distributed Brillouin optical time-domain analysis sensing system without relays, the optical fiber to be tested is 100 km long. The end 100m is wound down and placed in a temperature chamber. Then about 10m is wound out and placed on a stress plate to extract temperature and strain separately. After preprocessing, the experimental data can be input into the trained network model, such as Figure 3 As shown, the temperature and stress information along the optical fiber are finally obtained.
[0039] By applying the MPR-CNN network model described in this embodiment, the beneficial effects are as follows: When processing high-signal-to-noise ratio (SNR) BOTDA data after noise reduction by the MPR module, a convolutional neural network (CNN) is used to improve the accuracy and efficiency of data analysis. Traditional methods for extracting temperature and strain often rely on Lorentz curve fitting (LCF) at each sensing point along the fiber. As a distributed fiber-optic sensing technology, BOTDA generates massive amounts of data when measuring temperature and strain over long fiber lengths. LCF is extremely time-consuming, with processing 10,000 BGS data points taking approximately three minutes. Using a CNN to establish a mapping between input BGS data and output temperature and stress eliminates the need for curve fitting, increasing BOTDA measurement speed by several orders of magnitude. Furthermore, CNN models can automatically learn features from the data, eliminating the need for manual feature extraction. This is a significant advantage in many regression problems, especially when the data has complex and high-dimensional features. Furthermore, because convolutional layers can capture the spatial hierarchy and local connectivity patterns of images or data, CNNs excel at processing data with spatial correlation. Secondly, with the development of deep learning frameworks, CNNs can effectively process large-scale data sets, which is particularly important for parameter estimation and model training in regression tasks. In summary, applying MPR-CNN to BOTDA data processing can not only improve the accuracy of data analysis, but also accelerate the response capabilities of real-time monitoring and early warning systems, providing more reliable data support for key areas such as long-distance and large-scale structural health monitoring, earthquake prediction, and energy pipeline safety.
[0040] The MPR-CNN network, leveraging the powerful MPRnet denoising module, effectively processes BGS data with low or very low signal-to-noise ratios in multiple stages. MPRNet employs a multi-stage architecture, enabling progressively finer processing of BGS images, with each stage focusing on recovering distinct image features, resulting in more effective noise removal. By combining an encoder-decoder architecture with high-resolution branches, MPRNet is able to preserve key spatial details in BGS images while recovering rich contextual information. Its pixel-by-pixel adaptive design adjusts the denoising strength based on the characteristics of each pixel in the BGS image, helping to remove noise while maintaining signal integrity. MPRNet's multi-scale and multi-stage processing during training enhances the model's robustness to diverse noise types, which is particularly important for real-world BGS images. The CNN module, serving as the sensor information extraction network, is fast and accurate after training, capable of processing temperature and stress extraction from tens of thousands of BGS data points in seconds. It also exhibits strong generalization and can effectively handle data variations across diverse environments and conditions, resulting in enhanced robustness and reliability in practical applications. The CNN network module typically performs exceptionally well in regression tasks such as temperature and stress extraction. Thanks to its parallel processing capabilities, efficient feature extraction, optimized network structure, and automated parameter adjustment, it is typically faster than traditional methods or other models when performing curve fitting, significantly reducing computation time while maintaining high accuracy. Overall, the use of these two network modules enables the BOTDA system to achieve fast and precise responses.
[0041] It should be pointed out that, according to the needs of implementation, the various steps / components described in this application can be split into more steps / components, or two or more steps / components or partial operations of steps / components can be combined into new steps / components to achieve the purpose of the present invention.
[0042] It will be easily understood by those skilled in the art that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.
Claims
1. A temperature stress extraction method based on MPR-CNN, characterized in that: include: S1. Build and train the MPR-CNN network model based on the sensor device; S2. Obtain BGS images through sensing devices and input them into the MPR-CNN network model to obtain temperature and stress information.
2. The temperature stress extraction method based on MPR-CNN according to claim 1, characterized in that: The MPR-CNN network model includes MPR module and CNN module; The MPR module consists of an encoder-decoder subnetwork and an original resolution processing layer. The encoder-decoder subnetwork extracts the global context information of the input BGS image, and the original resolution processing layer performs convolution on the original resolution of the input BGS image to restore the fine details of the input BGS image. The encoder-decoder subnetwork is horizontally connected to the original resolution processing layer. The MPR module fuses the global context information and fine details to output a denoised BGS image.
3. The temperature stress extraction method based on MPR-CNN according to claim 2, characterized in that: The CNN module consists of an input layer, a convolutional layer, a pooling layer, and an output layer. The input layer is a one-dimensional vector whose size matches the number of BGS sweeps in the BGS image. The number of convolutional layers, the size of the convolution kernel, and the size of the pooling layer are all set based on the complexity of the Brillouin gain spectrum of a single sensor point in the BGS image. The output layer uses no activation function or a linear activation function.
4. The temperature stress extraction method based on MPR-CNN according to claim 2, characterized in that: The training process of the MPR module includes MPR training data construction and MPR training optimization. The MPR training data construction includes: constructing three-dimensional BGS training data with scanning frequency, fiber sensing distance, and signal amplitude as the X, Y, and Z axes respectively; converting the three-dimensional BGS training data into grayscale images in the XY two-dimensional plane; adding Gaussian white noise to the grayscale image to generate noisy training data with different signal-to-noise ratios; and matching the noisy training data with corresponding noise-free grayscale images as true labels.
5. The temperature stress extraction method based on MPR-CNN according to claim 4, characterized in that: The MPR module training optimization includes: inputting noisy training data into the MPR module; calculating the loss based on the denoised BGS image output by the MPR module and the true label of the noisy training data; backpropagating to update the MPR module parameters, and iterating until the loss converges.
6. The temperature stress extraction method based on MPR-CNN according to claim 4, characterized in that: The loss function used in the loss calculation of the MPR module is the mean square error or the structural similarity index.
7. The temperature stress extraction method based on MPR-CNN according to claim 3, characterized in that: The training process of the CNN module includes CNN training data construction and CNN training optimization; The CNN training data construction includes: using Brillouin frequency shift as a training label, and generating CNN training data consisting of multiple BGS images according to the set Brillouin frequency shift variation range, Brillouin linewidth variation, and signal-to-noise ratio; Normalize or standardize the CNN training data so that the CNN training data conforms to the standard normal distribution.
8. The temperature stress extraction method based on MPR-CNN according to claim 7, characterized in that: CNN training optimization includes: inputting CNN training data into the CNN module; calculating the loss based on the temperature and stress prediction information output by the CNN module and the training labels of the CNN training data; backpropagation to update the CNN module parameters, and iterating until the loss converges.
9. The temperature stress extraction method based on MPR-CNN according to claim 8, characterized in that: CNN training optimization includes: using the previous CNN module as the starting point of the current CNN training module for transfer learning; and applying the previous CNN module to the previous sensor system.
10. A temperature stress extraction system based on MPR-CNN, characterized in that: include: A sensing device, used for detecting and acquiring a BGS image; The MPR-CNN network model is used to obtain temperature and stress information based on BGS images.